Zero-Residual Concept Erasure via Progressive Alignment in Text-to-Image Model

Bowei Zhu; Hongxu Chen; Lin Li; Long Chen; Runshi Li; Taoran Mei; Zhen Wang

REVIEW 3 major objections 4 minor 20 references

Zero-Residual Concept Erasure via Progressive Alignment in Text-to-Image Model

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read ErasePro erases target concepts from text-to-image models by enforcing exactly zero alignment residual in a closed-form update, then spreading the edit from shallow to deep layers so quality-critical deep weights change least.

desk verdict ErasePro is a clean but incremental closed-form concept erasure method whose 'zero-residual' title oversells the property: exactness holds on the optimization features, not on unseen prompts, and the paper leaves the generalization claim untested. read the letter →

arxiv 2508.04472 v1 pith:KJR5KY3U submitted 2025-08-06 cs.CV cs.AIcs.LG

Hongxu Chen , Zhen Wang , Taoran Mei , Lin Li , Bowei Zhu , Runshi Li , Long Chen This is my paper

classification cs.CVcs.AIcs.LG

keywords concepterasuretext-to-imagediffusionclosed-formoptimizationzero-residualalignmentprogressivelayer-wiseupdateStablegenerativequalitypreservationsafetyfine-tuning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that existing closed-form concept-erasure methods — which solve an unconstrained least-squares alignment and apply the solution in one step to a few deep layers — leave a non-zero 'alignment residual', so target concepts survive in complex prompts, and they concentrate large parameter changes in the layers most tied to generative quality. ErasePro replaces the unconstrained objective with a constrained one: minimize the parameter deviation subject to $Wx_i = W_o y_i$ exactly, whose closed-form solution $W^{*} = W_o + (W_oY - W_oX)(X^{\top}X)^{-1}X^{\top}$ drives the residual to zero on the concept features. It then updates layers one at a time from shallow to deep, re-extracting features at each stage, so the target-to-anchor transition is gradual and deep layers need only small changes. The reported result is that ErasePro reaches CLIP accuracy 0.000 on the target concepts for instance, art-style, and nudity erasure while keeping anchor and unrelated-concept quality close to the original model. A reader should care because exact alignment on the optimization features plus small deep-layer deviations is precisely the combination the paper argues makes erasure both complete and non-destructive.

What carries the argument

Two devices carry the argument. First, the constrained closed-form solution $W^{*} = W_o + (W_oY - W_oX)(X^{\top}X)^{-1}X^{\top}$ — the minimizer of $\lVert W - W_o\rVert_F^2$ under the hard constraints $Wx_i = W_o y_i$ — drives the alignment residual $\lVert WX - W_oY\rVert_F^2$ to exactly zero on the concept features, unlike the unconstrained solution of prior methods. Second, the progressive protocol: after each layer is updated, inference re-extracts the next layer's input features, so targets are already partly anchored and deeper layers need only small deviations; an experiment injecting equal-magnitude synthetic deviations locates deep layers as the quality-critical ones. The solution

What would settle it

Take the erased model and several hundred held-out complex prompts containing the target concept (attributes, compositions, and contexts absent from the optimization set); extract their features $X'$ at each edited layer and compute $\lVert W^{*}X' - W_oY'\rVert_F$, and also generate images for an automatic detector (NudeNet for nudity, CLIP retrieval for general concepts). Any held-out prompt with a large residual or detectable target content shows the zero-residual guarantee does not generalize beyond the optimization set.

Watch

Extended reading notes

Core claim

ErasePro's central claim is that complete concept erasure requires changing the optimization objective, not just its solver. Existing closed-form methods minimize $\lVert WX - W_oY\rVert_F^2 + \lVert W - W_o\rVert_F^2$, whose solution still leaves a non-zero alignment residual; ErasePro instead solves $\min_W \lVert W - W_o\rVert_F^2$ subject to $Wx_i = W_o y_i$, giving the exact update $W^{*} = W_o + (W_oY - W_oX)(X^{\top}X)^{-1}X^{\top}$. Applied layer by layer from shallow to deep, re-extracting features at each stage, the required parameter deviation shrinks with depth; across instance, art-style, and nudity tasks the method reports CLIP accuracy 0.000 on all target concepts.

Load-bearing premise

The load-bearing premise is coverage: the finite set of prompt-derived concept features used in optimization must span how the target concept appears in complex unseen prompts — if it does not, the exactly-zero alignment guarantee holds only on the optimization set, and erasure stays incomplete elsewhere.

Editorial extensions

If this is right

On the features used during optimization, the alignment residual $\lVert W^{*}X - W_oY\rVert_F^2$ is exactly zero by construction, so erasure is exact for those prompts rather than approximate.
Because features are progressively aligned before each deep layer is edited, required parameter deviations shrink with depth, which is why generative quality on benign content stays close to the base model.
In the reported experiments ErasePro is the only method reaching CLIP accuracy 0.000 on every target (instance, art style, explicit nudity, implicit nudity), while UCE and AC retain measurable target signal (accuracy up to 0.287).
The whole procedure is closed-form and gradient-free, preserving the speed advantage of closed-form erasure while adding only repeated inference passes to re-extract features.
The same single procedure handles simultaneous erasure of multiple concepts (e.g., Trump, Pikachu, Minions, Garfield), so the constraint and progressive scheme scale to multi-target settings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The zero-residual guarantee is exact only for the optimization features, and the paper itself notes it holds at the objective level, not the full module (residual connections make the module-level residual non-zero). Coverage of unseen complex prompts is assumed, not analyzed; a stress test is to measure $\lVert W^{*}X' - W_oY'\rVert_F$ on held-out prompts.
The depth-sensitivity premise rests on one qualitative experiment with a synthetic identity-scaled deviation at $\alpha = 0.2$; if equal-norm real updates show a different shallow/deep ordering, the quality-preservation argument would need revisiting.
The same constrained-alignment identity applies to any linear layer (FFNs, convolutions, other projections), so widening the edited layer set beyond QKV projections is a natural, cheap way to broaden prompt coverage — an extension the paper does not explore.
Since each layer independently minimizes its own deviation, the scheme does not minimize the total edit across the network; a joint or sequentially regularized variant could trade per-layer exactness for an even smaller cumulative parameter change.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The core of the paper is simple and correct: instead of the regularized least-squares objective used by UCE, they solve the constrained least-norm problem W* = argmin ||W-Wo||² s.t. WX = WoY, which gives the standard closed-form W* = Wo + (WoY - WoX)(XᵀX)⁻¹Xᵀ. That is the same least-norm solution used in ROME/MEMIT-style model editing, and the paper should have cited that line of work; the novelty is the layer-by-layer progressive application from the text encoder into the U-Net cross-attention, and the empirical demonstration on instance, style, and nudity erasure. The appendix tables are reasonably thorough, and the method does seem to erase better than UCE, AC, and ESD on the tested tasks, with a sensible quality/trade-off knob for the implicit nudity variant. So there is real incremental value here, and the math checks out under the stated full-column-rank assumption. The soft spots are in the claims, not the computations. The 'zero-residual' result is true only for the finite set of features X used in the optimization. Footnote 3 already concedes the full-module residual is nonzero due to residual connections, and the stress-test derivation makes the stronger point: for a new prompt feature x', even if x' lies in span(X), the residual is W_o(Yc - y') where c is the least-squares coefficient of x' on X. Nothing in the method guarantees y' is the same linear combination of anchor features, so 'perfect alignment' is not a statement about generation at inference time. The experiments use only 10 explicit prompts per concept, so the paper's motivating story about complex prompts amplifying misalignment is asserted, not empirically covered. The progressive scheme's core assumption—that deeper layers are more sensitive—rests on one qualitative study (Appendix A.2, alpha=0.2), and the implicit nudity variant introduces two free hyperparameters (token count and starting layer) without a sensitivity sweep. There are also no error bars and no code release, which makes the headline numbers hard to judge. Who is this for? People working on safety/editing in diffusion models will find it a useful incremental baseline. It deserves a serious referee, but not because the title's promise of 'zero-residual erasure' holds up—it does not. It deserves review because the constrained-optimization-plus-progressive scheme is a clean, reproducible idea that can be made honest with a corrected narrative, a ROME/MEMIT citation, and held-out-prompt experiments that actually vary prompt complexity. Send it to review, but tell the authors to rewrite the title and abstract and to substantiate the generalization claim empirically.

Referee Report

3 major / 4 minor

Summary. The paper proposes ErasePro, a closed-form concept-erasure method for text-to-image models. It introduces a constrained objective (Eq. 4) whose solution (Eq. 5) maps a finite set of target features exactly onto anchor features, and a progressive, layer-wise update procedure that modifies text-encoder and cross-attention layers from shallow to deep. Experiments on instance, art-style, and explicit/implicit nudity erasure report strong quantitative results on CLIP/KID/FID and NudeNet, including perfect CLIP accuracy on the tested targets. The mathematical derivation is standard under a full-column-rank assumption; the contribution is the constrained formulation and the progressive update scheme.

Significance. If the generalization step held, ErasePro would be an efficient and simple alternative to gradient-based erasure, with the unusual property of exactly zero alignment residual on the optimization set. The appendix derivation of Eq. 5 is correct, and the evaluation is broader than typical (I2P/NudeNet, COCO 30k, multi-concept). The paper also clearly identifies the two limitations it addresses. However, the central 'zero-residual' and quality-preservation claims are stronger than what is proven: the residual is zero only on the finite feature set and within the linear objective, and the progressive scheme is not isolated by ablation. The empirical results remain valuable, but the conceptual claims need to be scoped or supported.

major comments (3)

[§3.2, Eq. (5)] Eq. 5 guarantees WX = WoY only on the finite feature set X, not on unseen prompts. For a held-out pair (x', y'), W*x' - Wo y' = Wo[(x' - Xc)+(Yc - y')], c=(X^T X)^{-1}X^T x'; even if x' lies in span(X), the residual is Wo(Yc-y') and is not zero unless y' is the same linear combination of anchor features. The sentence 'ensures a full alignment ... achieving more complete erasure for target concepts' in §3.2 is therefore an unsupported generalization. The per-concept evaluation uses only 10 prompts (Appendix A.2), so coverage of complex prompts is not measured. Please add residual/coverage experiments or scope the zero-residual claim to the optimization objective.
[§3.2, Improvement 2; Appendix A.2] The quality-preservation claim in Improvement 2 rests on the assumption that deeper layers are more sensitive to parameter changes, but the supporting evidence is one qualitative experiment at α=0.2 (Appendix A.2, Figure 5). More importantly, there is no ablation that isolates the progressive update scheme from the constrained objective: no variant applies Eq. 5 only to deep layers, or updates the same layer set non-progressively. Without such a control, Tables 1-4 cannot attribute the quality gains to progressivity rather than to the different solver. Please add a controlled comparison and, if feasible, quantitative sensitivity measurements over several α values.
[Footnote 3 / Abstract] Footnote 3 concedes that the full-module residual is inherently non-zero due to residual connections and that the zero-residual claim refers only to the objective in Eq. 4. However, the title and abstract advertise 'Zero-Residual Concept Erasure' and a 'strict zero-residual constraint'. Please align the terminology with the actual scope of the guarantee.

minor comments (4)

[§4.1] The text states 'five main baselines' but then lists AC, ESD-x, ESD-u, and UCE; SDv2.1 appears later for nudity. Please clarify the count or add the missing baseline.
[Appendix A.2, Algorithm 2] Algorithm 2 computes W* but appears to omit the 'Layer_i ← Update Weights(Layer_i, W*)' step that is present in Algorithm 1. As written, the updated weights are never applied inside the loop.
[§4.2] There are duplicated/placeholder references: 'Figure ??' appears for nudity results, and the paragraph beginning 'we generated 4,703 images using prompts...' is repeated verbatim in §4.1 and §4.2. Please clean up.
[§4.2, nudity erasure] The empty-string anchor for nudity erasure needs a precise description of how its feature embedding is constructed, since that choice is not obvious and presumably affects the optimization.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: zero residual is an enforced constraint, not a fitted prediction; the only self-citation is non-load-bearing.

full rationale

The central 'zero-residual' claim is not circular in the problematic sense. Eq. 5 is derived as the KKT solution to the constrained problem Eq. 4, and the property W*X = W_oY follows immediately from the constraint WX = W_oY that Eq. 4 imposes. The paper does not fit a parameter and then relabel the fit as a prediction; it transparently constructs a constraint whose solution has zero residual on the optimization features. The empirical contribution is measured downstream: CLIP accuracy and KID on generated images per concept, NudeNet statistics on 4,703 I2P prompts, and FID/KID on COCO 30k, all independent of the closed-form solution. The only self-citation is to Iteris (Chen et al. 2025) in a related-work list on personalized customization; it is not invoked to justify Eq. 4, Eq. 5, the progressive alignment, or any empirical claim, so it is not load-bearing. The paper's own footnote ('From a full-module perspective, it is inherently non-zero due to residual connections') correctly limits the zero-residual claim to the objective, and the extension to unseen complex prompts is a coverage assumption rather than a proved result; that is a correctness/overclaim concern, not circularity. Score is 2 only because of the minor non-load-bearing self-citation.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the assumptions that the features used in optimization are representative, that projection-level alignment suffices for concept erasure, and that deeper layers are more sensitive to edits. These are mostly domain assumptions with partial empirical support. The only hand-chosen numbers are the implicit-erasure variants' token counts and starting layers.

free parameters (3)

Implicit-erasure token count = 5 (ErasePro-w), 8 (ErasePro-s)
Hand-chosen per variant; appendix A.2 states these values, affecting the erasure strength and quality trade-off.
Implicit-erasure starting layer = 5th (w), 1st (s)
Hand-chosen; the appendix says intervention begins at the 5th layer for the weak variant and the 1st layer for the strong variant.
Layer count S for explicit tasks = Not specified
The algorithm input S is never given a concrete value or search procedure in the main text or appendix.

assumptions (4)

domain assumption The target feature matrix X is column full rank, so (X^T X)^-1 exists.
Required for Eq. 5; the appendix acknowledges this and suggests the pseudoinverse otherwise. Likely satisfied in practice but not guaranteed.
domain assumption Layer-wise erasure can be achieved by updating only linear projection matrices (QKV projections) within attention layers.
The objective aligns the output of a projection W applied to features, but actual layers contain nonlinear attention and MLP operations; the sufficiency of projection-level alignment is assumed.
domain assumption Deeper layers are more sensitive to parameter deviations than shallow layers.
The progressive design relies on this sensitivity ordering. Support is a single qualitative experiment with alpha=0.2 in Appendix A.2, not a measured robustness curve.
domain assumption Zero-residual alignment on a finite set of concept features transfers to unseen prompts containing the target concept.
This is the core generalization premise of the method; the paper motivates it via the failure mode of prior methods but does not prove or measure feature coverage.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Zero-Residual Concept Erasure via Progressive Alignment in Text-to-Image Model." pith.science (2026). https://pith.science/paper/KJR5KY3U

@misc{pith2026250804472,
  author       = {Pith},
  title        = {Pith review of: Zero-Residual Concept Erasure via Progressive Alignment in Text-to-Image Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KJR5KY3U}},
  note         = {Machine review of arXiv:2508.04472}
}

read the original abstract

Concept Erasure, which aims to prevent pretrained text-to-image models from generating content associated with semantic-harmful concepts (i.e., target concepts), is getting increased attention. State-of-the-art methods formulate this task as an optimization problem: they align all target concepts with semantic-harmless anchor concepts, and apply closed-form solutions to update the model accordingly. While these closed-form methods are efficient, we argue that existing methods have two overlooked limitations: 1) They often result in incomplete erasure due to "non-zero alignment residual", especially when text prompts are relatively complex. 2) They may suffer from generation quality degradation as they always concentrate parameter updates in a few deep layers. To address these issues, we propose a novel closed-form method ErasePro: it is designed for more complete concept erasure and better preserving overall generative quality. Specifically, ErasePro first introduces a strict zero-residual constraint into the optimization objective, ensuring perfect alignment between target and anchor concept features and enabling more complete erasure. Secondly, it employs a progressive, layer-wise update strategy that gradually transfers target concept features to those of the anchor concept from shallow to deep layers. As the depth increases, the required parameter changes diminish, thereby reducing deviations in sensitive deep layers and preserving generative quality. Empirical results across different concept erasure tasks (including instance, art style, and nudity erasure) have demonstrated the effectiveness of our ErasePro.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 9 canonical work pages

[1]

inappropriateness

The Inappropriate Image Prompts (I2P) dataset Schramowski et al. (2023) is a benchmark designed to evaluate the tendency of text-to-image diffusion models to generate inappropriate content. It contains real-world user prompts that are disproportionately likely to produce harmful or sensitive imagery, including hate, harassment, violence, self-harm, sexual...

work page 2023
[2]

explicitly

In erasure tasks, where the target concepts “explicitly” appear in the prompt, the generation for each concept was conditioned on 10 prompts. All prompts were generated using GPT-4o Hurst et al. (2024) to ensure they are well-suited for T2I models. Inference Setting. We adopted stable diffusion v1.4 as base model. During inference, we used the following c...

work page 2024
[3]

Gpt-4o system card

Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Os- trow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276,

arXiv
[4]

Ablating concepts in text-to-image diffusion models

Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In ICCV, pp. 22691–22702, 2023a. Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, pp. 1931–1941, 2023b. Ouxiang Li, Y...

arXiv 1931
[7]

Hierarchical text- conditional image generation with clip latents

Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text- conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3,

arXiv
[8]

Red-teaming the stable diffusion safety filter

Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tram `er. Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610,

arXiv
[9]

Laion-400m: Open dataset of clip-filtered 400 million image-text pairs

Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114,

arXiv
[10]

Ai art generators hit with copyright suit over artists’ images.Bloomberg Law

Riddhi Setty. Ai art generators hit with copyright suit over artists’ images.Bloomberg Law. Accessed on February, 1:2023,

work page 2023

Show all 20 references

[11]

Precise parameter localization for textual generation in diffusion models

Łukasz Staniszewski, Bartosz Cywi ´nski, Franziska Boenisch, Kamil Deja, and Adam Dziedzic. Precise parameter localization for textual generation in diffusion models. arXiv preprint arXiv:2502.09935,
[12]

Safree: Training-free and adaptive guard for safe text-to-image and video generation

Jaehong Yoon, Shoubin Yu, Vaidehi Patil, Huaxiu Yao, and Mohit Bansal. Safree: Training-free and adaptive guard for safe text-to-image and video generation. arXiv preprint arXiv:2410.12761 ,

arXiv
[14]

3 is non-zero in general, i.e., for almost all X ̸= Y

We aim to show that Eq. 3 is non-zero in general, i.e., for almost all X ̸= Y. Let us denote the inner term by: ∆ = (YX⊤ + I)(XX⊤ + I)−1X − Y. If X ̸= Y, then ∆ ̸= 0 in almost all cases (except for rare degenerate cases), because: • When X = Y, it is easy to verify that ∆ = 0,...

2022
[17]

(b) More qualitative results for implicit nudity erasure

For the weak variant (ErasePro-w), concept features are built by concatenating the first 13 �� Figure 6: (a) More qualitative results for explicit nudity erasur...

2024
[18]

We used the official diffusers implementation of Ablating Concepts (AC) Kumari et al

Training Setting of AC. We used the official diffusers implementation of Ablating Concepts (AC) Kumari et al. (2023a) under default settings. For instance erasure, we fine-tune the model by updating cross-attention layers with a learning rate of 2e-6, batch size 4, and 100 tra...

2023
[19]

(2018), anchor CLIP Radford et al

Our method consistently achieves the best performance across almost all cases in terms of anchor KID Bi ´nkowski et al. (2018), anchor CLIP Radford et al. (2021) score, and CLIP accuracy. Moreover, it generally outperforms other baselines in preserving the integrity of other c...

2018
[20]

Pikachu”, “Minions

In this setting, we simultaneously erase the concepts of “Trump”, 14 �� ...

1900
[2021]

P- tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks

Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Lam Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. P- tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602,

arXiv
[2022]

Demystifying mmd gans

Mikołaj Bi´nkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401,

arXiv
[2023]

Muse: Text-to-image gen- eration via masked generative transformers

Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image gen- eration via masked generative transformers. arXiv preprint arXiv:2301.00704,

arXiv
[2024]

Scaling autoregressive models for content- rich text-to-image generation

Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content- rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5,

arXiv
[2025]

Prefix-tuning: Optimizing continuous prompts for generation.arXiv preprint arXiv:2101.00190,

Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation.arXiv preprint arXiv:2101.00190,

arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

inappropriateness

The Inappropriate Image Prompts (I2P) dataset Schramowski et al. (2023) is a benchmark designed to evaluate the tendency of text-to-image diffusion models to generate inappropriate content. It contains real-world user prompts that are disproportionately likely to produce harmful or sensitive imagery, including hate, harassment, violence, self-harm, sexual...

work page 2023

[2] [2]

explicitly

In erasure tasks, where the target concepts “explicitly” appear in the prompt, the generation for each concept was conditioned on 10 prompts. All prompts were generated using GPT-4o Hurst et al. (2024) to ensure they are well-suited for T2I models. Inference Setting. We adopted stable diffusion v1.4 as base model. During inference, we used the following c...

work page 2024

[3] [3]

Gpt-4o system card

Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Os- trow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276,

arXiv

[4] [4]

Ablating concepts in text-to-image diffusion models

Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In ICCV, pp. 22691–22702, 2023a. Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR, pp. 1931–1941, 2023b. Ouxiang Li, Y...

arXiv 1931

[5] [7]

Hierarchical text- conditional image generation with clip latents

Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text- conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3,

arXiv

[6] [8]

Red-teaming the stable diffusion safety filter

Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tram `er. Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610,

arXiv

[7] [9]

Laion-400m: Open dataset of clip-filtered 400 million image-text pairs

Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114,

arXiv

[8] [10]

Ai art generators hit with copyright suit over artists’ images.Bloomberg Law

Riddhi Setty. Ai art generators hit with copyright suit over artists’ images.Bloomberg Law. Accessed on February, 1:2023,

work page 2023